Energy treatment method for mixed kitchen garbage

Abstract

The invention discloses a mixed kitchen garbage energy treatment method, and relates to the technical field of energy treatment, and the method comprises the steps: collecting initial component data of mixed kitchen garbage, encrypting the initial component data, outputting regional encrypted data, uploading the regional encrypted data to a federal learning model training platform, aggregating the multi-regional encrypted data, training a global garbage component prediction model, and carrying out the energy treatment of the mixed kitchen garbage. Outputting a component distribution probability matrix of the mixed kitchen garbage in a future period; the component distribution probability matrix is input into a hyperspectral sorting machine, spatial distribution of organic matter, grease and impurities in unsorted mixed garbage flow is analyzed in real time through 3D-CNN, and a dynamic sorting instruction set is generated; and tracking energy output data and carbon emission intensity of the whole process through a digital thread, dynamically correcting a weight coefficient of the federated learning model, and updating an energy conversion rule base. According to the method, the federal learning model and the hyperspectral sorting technology are combined, the spatial and temporal distribution characteristics of the mixed kitchen garbage components are accurately captured, and the sorting precision and the energy conversion efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of energy treatment, and in particular to a method for energy treatment of mixed food waste. Background Art

[0002] With the acceleration of the urbanization process and the continuous expansion of the population scale, food waste has become the component with the highest proportion in urban solid waste. Its efficient treatment and energy utilization are of great significance for alleviating environmental pressure, promoting the circular economy of resources, and achieving the dual-carbon strategic goal. Food waste is rich in organic matter, and high-value energy products such as biogas and biohydrogen can be generated through biological conversion technology, which is one of the core ways to replace fossil fuels and reduce carbon emissions. Therefore, the development of efficient and stable energy treatment technologies for mixed food waste has become a key research direction in the field of waste resource utilization.

[0003] Traditional energy treatment methods for food waste are mainly based on a single process route, and energy conversion is achieved through technologies such as anaerobic fermentation or pyrolysis gasification. However, when facing mixed food waste with complex components and large physical property differences, especially in a complex system where multiple factors such as oil, fiber, and moisture interact, the energy conversion efficiency and operation stability of traditional processes are significantly reduced, seriously restricting the feasibility of large-scale application. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for energy treatment of mixed food waste to solve the problems of low conversion efficiency and poor operation stability of traditional single processes when treating food waste with complex components.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a method for the energy conversion treatment of mixed food waste, which includes collecting initial composition data of the mixed food waste, encrypting it, and then uploading the encrypted data of the region to a federated learning model training platform. Aggregate the encrypted data of multiple regions and train a global garbage composition prediction model to output a composition distribution probability matrix of the mixed food waste in the future period. Input the composition distribution probability matrix into a hyperspectral sorter, and use 3D-CNN to real-time analyze the spatial distribution of organic matter, oil, and impurities in the unsorted mixed waste stream to generate a dynamic sorting instruction set. Control a multi-stage linkage device according to the dynamic sorting instruction set to output a high-oil garbage stream and a high-fiber garbage stream with homogenized components. Match a preset energy conversion rule library through a reinforcement learning decision engine, dynamically select the optimal energy conversion path, and generate a process parameter instruction set. Send the process parameter instruction set to the energy conversion device, control the temperature gradient, stirring frequency, and bacterial inoculation amount of the anaerobic reactor, and adjust the catalyst concentration and reaction duration of the biodiesel synthesis reactor, while synchronously and real-time monitoring the methane production rate and biodiesel purity. Track the energy output data and carbon emission intensity of the entire process through a digital thread, dynamically correct the weight coefficients of the federated learning model, and update the energy conversion rule library.

[0007] As a preferred embodiment of the method for the energy conversion treatment of mixed food waste according to the present invention, wherein: the output of the composition distribution probability matrix of the mixed food waste in the future period includes the following steps, In the federated learning model training platform, perform homomorphic decryption on the encrypted data of each region, add differential privacy noise to the decrypted data, and extract local model parameters; Aggregate the local model parameters through a dynamic weight allocation algorithm, and jointly determine the weights of the local model parameters by the historical variation coefficient and seasonal fluctuation amplitude value of the regional garbage composition; Based on the weights of the local model parameters, construct a spatio-temporal feature fusion space through an attention mechanism, and integrate the time series gradient and spatial covariance matrix of the garbage composition in each region into a global spatio-temporal tensor; Input the global spatio-temporal tensor into a Bayesian probability network to generate a composition distribution probability matrix.

[0008] As a preferred embodiment of the method for the energy conversion treatment of mixed food waste according to the present invention, wherein: the generation of the dynamic sorting instruction set includes the following steps, Collect multi-band reflection spectral data of the unsorted mixed waste stream to construct a three-dimensional hyperspectral tensor; The unsorted mixed waste stream refers to the physical material that enters the hyperspectral sorter in real time through a conveyor belt; Use the depthwise separable convolutional layer of 3D-CNN to extract the spatial distribution characteristics of organic matter, oil, and impurities, and synchronously calculate the temporal change rate of the spectral absorption peak; Fusing the spatial distribution characteristics and the temporal change rate of spectral absorption peaks, predicting the sorting lag compensation amount through a gated recurrent unit, and generating a three-dimensional dynamic sorting instruction set.

[0009] As a preferred embodiment of the method for energy conversion of mixed food waste according to the present invention, wherein: the step of dynamically selecting the optimal energy conversion path and generating a process parameter instruction set includes the following steps: Real-time monitoring of the real-time methane production rate, biodiesel purity, catalyst activity coefficient, and reactor temperature gradient in the multi-stage linkage equipment as monitoring data, and constructing a reinforcement learning state space; Based on the reinforcement learning state space, constructing a multi-objective reward function by combining the energy output index and the carbon emission intensity index in a weighted summation manner; Based on the reinforcement learning state space and the multi-objective reward function, performing deep deterministic policy gradient algorithm optimization, and outputting the range of anaerobic digestion temperature adjustment, the increment of stirring frequency, and the compensation coefficient of the microbial inoculum dosage; Dynamically adjusting the weight ratio of energy output to carbon emission in the multi-objective reward function, and updating the parameter mapping relationship in the energy conversion rule base to generate a process parameter instruction set.

[0010] As a preferred embodiment of the method for energy conversion of mixed food waste according to the present invention, wherein: the initial ingredient data includes oil content, moisture ratio, and impurity type.

[0011] As a preferred embodiment of the method for energy conversion of mixed food waste according to the present invention, wherein: the hyperspectral sorter refers to a device integrating hyperspectral imaging, 3D-CNN real-time analysis, and sorting instruction issuing functions.

[0012] As a preferred embodiment of the method for energy conversion of mixed food waste according to the present invention, wherein: controlling the multi-stage linkage equipment includes adaptively adjusting the rotational speed gradient of the shear crusher, the pressure threshold of the screw dehydrator, and the triggering condition for performing low-temperature deoiling.

[0013] As a preferred embodiment of the method for energy conversion of mixed food waste according to the present invention, wherein: the high-oil food waste stream triggers a transesterification reaction to generate biodiesel, and the high-fiber food waste stream triggers a two-phase anaerobic digestion to generate biogas.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and wherein: when the computer program is executed by the processor, any step of the method for energy conversion of mixed food waste as described in the first aspect of the present invention is implemented.

[0015] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program thereon, wherein: when the computer program is executed by a processor, any step of the method for energy conversion treatment of mixed food waste as described in the first aspect of the present invention is implemented.

[0016] The beneficial effects of the present invention are as follows: By combining the federated learning model with the hyperspectral sorting technology, the spatio-temporal distribution characteristics of the components of mixed food waste are accurately captured, and the sorting accuracy and energy conversion efficiency are improved. By fusing multi-region encrypted component data with the differential privacy perturbation mechanism, a global spatio-temporal tensor prediction model is constructed to enhance the robustness and data privacy security of dynamic prediction of waste components. By adopting multi-objective carbon emission constraint conditions in the reinforcement learning decision engine, the optimization of the energy conversion process is deeply coupled with the carbon neutrality goal, enhancing the environmental friendliness and energy efficiency synergy of the treatment process, and realizing the efficient and low-carbon closed-loop operation of the energy conversion treatment of food waste. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0018] Figure 1 It is a flowchart of the method for energy conversion treatment of mixed food waste.

[0019] Figure 2 It is a flowchart of the federated learning model training.

[0020] Figure 3 It is a flowchart of the hyperspectral dynamic sorting control.

[0021] Figure 4 It is a flowchart of the reinforcement learning energy conversion decision. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be made in conjunction with the accompanying drawings of the specification.

[0023] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention, but the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the spirit of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0024] Second, the "one embodiment" or "embodiment" referred to herein means a specific feature, structure, or characteristic that may be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it an individual or selectively mutually exclusive embodiment with other embodiments.

[0025] Referring to Figures 1 to 4 , which is an embodiment of the present invention. This embodiment provides a method for energy conversion treatment of mixed food waste, including the following steps: S1: Collect the initial composition data of the mixed food waste, encrypt it, and then upload the regionally encrypted data to the federated learning model training platform. Aggregate the encrypted data from multiple regions and train the global garbage composition prediction model to output the composition distribution probability matrix of the mixed food waste in the future period.

[0026] Specifically, it includes the following steps: S1.1: Through a hyperspectral imaging sensor, collect the spectral data reflected from the surface of the garbage stream (such as the 400 - 2500 nm band), generate the original spectral image sequence, and analyze the spatial distribution information of the components on the surface of the garbage stream.

[0027] Specifically, based on the original spectral image sequence, use an image classification algorithm (based on the ResNet-50 pre-trained model) to identify the types of impurities and output the pixel distribution ratios of plastics, metals, and glass (types of impurities).

[0028] It should be noted that the spatial distribution information provided in the original spectral image sequence provides a spatial reference for the scanning position of the near-infrared spectrometer. Example: If the original spectral image sequence shows that there are metal impurities in a certain area, the near-infrared spectrometer can specifically enhance the scanning density of this area.

[0029] Furthermore, scan the inside of the garbage stream in real time through a near-infrared spectrometer to measure the oil content (such as the absorption peak intensity in the 900 - 1700 nm band) and the moisture ratio (such as the absorption peak area ratio of 1450 nm to 1940 nm).

[0030] It should be noted that the oil content and moisture ratio output by the near-infrared spectrometer are jointly calibrated with the surface composition data (spatial distribution information of the components on the surface of the garbage stream) in the hyperspectral image sequence to form the initial composition data of the mixed food waste. Among them, the initial composition data includes surface components (types of impurities) and internal components (oil content, moisture ratio).

[0031] S1.2: Perform the Paillier homomorphic encryption algorithm on the initial composition data to generate regionally encrypted data.

[0032] Encapsulate the regional encrypted data into JSON format and upload it to the federated learning model training platform through the HTTPS protocol.

[0033] S1.3: In the federated learning model training platform, perform homomorphic decryption on the encrypted data of each region, add differential privacy noise to the decrypted data, and extract local model parameters.

[0034] Specifically, in the federated learning model training platform, use the private key of the Paillier homomorphic encryption algorithm to perform decryption operations on the encrypted data of each region. After restoring it to the initial component data in plaintext format, add Laplace noise (noise scale parameter ε = 0.1) to each field of the decrypted data to generate a perturbed data set that satisfies differential privacy protection. Based on this perturbed data set, train a local LSTM model and extract the weight matrix and time gradient of its hidden layer as local model parameters.

[0035] Aggregate the local model parameters through a dynamic weight allocation algorithm, and jointly determine the weights of the local model parameters by the historical variation coefficient and seasonal fluctuation amplitude value of the regional garbage components.

[0036] Specifically, input the local model parameters of each region into the dynamic weight allocation algorithm. By calculating the historical variation coefficient of the garbage components in each region (such as the ratio of the standard deviation to the mean in the past three months) and the seasonal fluctuation amplitude value (such as the range of components in the same month in the past year), perform weighted calculation to determine the aggregation weights of the local model parameters.

[0037] It should be noted that the weighted calculation adaptively adjusts the contribution degrees of the parameters of each region by dynamically fusing the historical data variability and seasonal fluctuation characteristics. Specifically, the historical variation coefficient reflects the short-term fluctuation law of the regional garbage components, and the seasonal fluctuation amplitude value depicts the long-term periodic change trend. The weighted combination of the two can more accurately characterize the spatio-temporal characteristics of the regional data, thereby improving the robustness of the federated model aggregation. The proportion of weight allocation can be adjusted according to the actual scenario requirements or experimental optimization results.

[0038] Based on the weights of the local model parameters, construct a spatio-temporal feature fusion space through an attention mechanism, and integrate the time series gradient and spatial covariance matrix of the garbage components in each region into a global spatio-temporal tensor.

[0039] Specifically, based on the weighted local model parameters, use a multi-head attention mechanism (number of heads = 8) to construct a spatio-temporal feature fusion space, where the time dimension is associated with the time series gradient of the components in each region (step size = 24 hours), and the spatial dimension is associated with the covariance matrix between regions (similarity calculated based on geographical distance), and output the fused global spatio-temporal tensor.

[0040] Input the global spatio-temporal tensor into the Bayesian probability network to generate a component distribution probability matrix.

[0041] Specifically, the nodes of the Bayesian probability network include continuous nodes (oil content and moisture ratio) and discrete nodes (plastic, metal, and glass). By calculating the component distribution probability matrix of the mixed food waste within the Markov chain Monte Carlo sampling period (such as the probability distribution of each component within the next 24 hours), a component distribution probability matrix is generated (such as dimension = 24×5×1).

[0042] S2: Input the component distribution probability matrix into the hyperspectral sorter, and through 3D-CNN, the spatial distributions of organic matter, oil, and impurities in the unsorted mixed waste stream are analyzed in real time to generate a dynamic sorting instruction set.

[0043] Among them, the spatial distribution of impurities refers to the spatial distribution data such as the specific position coordinates, aggregation density, and coverage area ratio of impurities such as plastic, metal, and glass in the three-dimensional space of the conveyor belt.

[0044] Specifically, it includes the following steps: S2.1: Collect the multi-band reflection spectral data of the unsorted mixed waste stream and construct a three-dimensional hyperspectral tensor.

[0045] Specifically, during the process of the conveyor belt transmitting the unsorted mixed waste stream (physical material) to the hyperspectral sorter in real time, the reflection spectral data on the surface of the waste stream (such as the reflection spectral data in the 400 - 2500nm band) is collected through the hyperspectral imaging module built into the sorter to generate a three-dimensional hyperspectral tensor.

[0046] It should be noted that the dimensions of the three-dimensional hyperspectral tensor include spatial height, spatial width, and the number of spectral bands; among them, the spatial height and width correspond to the resolution of the real-time image of the conveyor belt.

[0047] S2.2: Use the depthwise separable convolutional layer of 3D-CNN to extract the spatial distribution feature maps of organic matter, oil, and impurities, and synchronously calculate the temporal change rate of the spectral absorption peak.

[0048] Specifically, input the three-dimensional hyperspectral tensor into the depthwise separable convolutional layer of the 3D-CNN (convolution kernel size 3×3×5, stride 1×1×2), and layer by layer extract the spatial distribution feature maps of organic matter, oil, and impurities; synchronously calculate the temporal change rate of the spectral absorption peak, and based on the intensity difference of the absorption peaks in the same spectral band of adjacent time frames (such as with an interval of 1 second), generate an absorption peak change rate matrix.

[0049] It should be noted that the spatial distribution feature map refers to the multi-dimensional tensor data output after processing the impurity spatial distribution data through the depthwise separable convolutional layer of 3D-CNN, which represents the distribution pattern of organic matter, oil, and impurities in the three-dimensional space of the conveyor belt in a rasterized form.

[0050] S2.3: Integrate the spatial distribution features and the temporal change rate of spectral absorption peaks, and predict the sorting lag compensation amount through a gated recurrent unit.

[0051] Specifically, concatenate the spatial distribution feature map and the absorption peak change rate matrix along the channel dimension, and input them into a gated recurrent unit (GRU, number of hidden layer units = 64). Predict the execution lag time of the hyperspectral sorting robotic arm (such as 0.2 seconds) through time series rolling prediction, and generate a sorting lag compensation amount matrix. Based on the lag compensation amount matrix, dynamically adjust the trigger time point of the sorting instruction to generate an immediate sorting instruction set, where the types of immediate sorting instructions include shearing and crushing, air jet sorting, and low-temperature deoiling.

[0052] S2.4: Input the component distribution probability matrix into the hyperspectral sorter to generate a three-dimensional dynamic sorting instruction set.

[0053] Specifically, encode the component distribution probability matrix output by the federated learning model training platform into a binary stream and transmit it in real time through the data interface of the hyperspectral sorter.

[0054] After the hyperspectral sorter parses the binary stream, align the component distribution probability matrix and the immediate sorting instruction set in space and time (such as mapping the future 1-hour probability distribution to the current sorting action) to generate a three-dimensional dynamic sorting instruction set.

[0055] Furthermore, at the end of the conveyor belt, sample and detect the actual sorted waste component distribution (such as grease residue rate, impurity ratio), calculate the KL divergence (Kullback-Leibler divergence) between the actual distribution and the probability matrix. If the divergence value > 0.1 (example), trigger the weight correction process of the federated learning model training platform to update the local model parameter weights and the spatio-temporal covariance matrix. Otherwise, keep the current federated learning model parameters unchanged and continue to execute the component prediction and three-dimensional dynamic sorting instruction set generation for the next cycle.

[0056] Preferably, the three-dimensional dynamic sorting instruction set dynamically optimizes the sorting action parameters (such as air jet sorting pressure) by integrating real-time detection data and future component predictions to ensure high-purity component separation, providing homogenized raw materials for subsequent anaerobic digestion and biodiesel synthesis. At the same time, feedback the execution effect to the federated learning model through the KL divergence detection at the end of the conveyor belt to form a closed-loop optimization, directly control the multi-stage linkage equipment to complete the separation of the waste flow components, and trigger the energy conversion paths of the high-grease waste flow and the high-fiber waste flow. Finally, track the overall process energy efficiency and carbon emission intensity through a digital thread.

[0057] S3: Control the multi - level linkage equipment according to the dynamic sorting instruction set, output high - grease waste streams and high - fiber waste streams with homogenized components, and through the reinforcement learning decision - making engine, match the preset energy conversion rule base, dynamically select the optimal energy conversion path and generate a process parameter instruction set.

[0058] Specifically, it includes the following steps: S3.1: Real - time monitor the real - time methane production rate, biodiesel purity, catalyst activity coefficient and reactor temperature gradient in the multi - level linkage equipment as monitoring data, and construct a reinforcement learning state space.

[0059] Specifically, based on the dynamic sorting instruction set, control the rotational speed gradient of the shearing crusher, the pressure threshold of the screw dehydrator and the low - temperature oil - removal trigger condition of the multi - level linkage equipment, and output high - grease waste streams and high - fiber waste streams with homogenized components; measure the methane production rate in real - time through the gas flowmeter of the anaerobic reactor, obtain the biodiesel purity through the spectral detector of the biodiesel synthesis reactor, collect the catalyst activity coefficient using an electrochemical sensor, and record the reactor temperature gradient through a temperature sensor, and construct a four - dimensional reinforcement learning state space.

[0060] It should be noted that the rotational speed gradient of the shearing crusher, the pressure threshold of the screw dehydrator and the low - temperature oil - removal trigger condition of the multi - level linkage equipment are derived from the real - time analysis and prediction optimization of the composition of the mixed waste stream by the dynamic sorting instruction set. Specifically, the three - dimensional dynamic sorting instruction set analyzed by the hyperspectral sorter is used to obtain the three - dimensional spatial distribution characteristics of grease, fiber and impurities in the waste stream, and is dynamically set based on the spatial coordinates and density of different component aggregation regions.

[0061] For example, the rotational speed gradient of the shearing crusher is based on the spatial coverage ratio of the high - fiber waste stream region (when the proportion of the region with fiber content ≥ 60% reaches 30%, the rotational speed is increased to 800 rpm), the pressure threshold of the screw dehydrator is based on the grease concentration distribution in the high - grease waste stream region (when the grease content ≥ 35%, the pressure threshold is adjusted to 1.5 MPa), and the low - temperature oil - removal trigger condition is through the real - time grease absorption peak intensity threshold (when the absorption peak area ratio in the 900 - 1700 nm band > 0.8, trigger low - temperature oil - removal at - 5°C).

[0062] S3.2: Based on the reinforcement learning state space, construct a multi - objective reward function by combining the energy output index and the carbon emission intensity index in a weighted summation manner.

[0063] Specifically, match the four - dimensional reinforcement learning state space data (methane production rate, biodiesel purity, catalyst activity coefficient, temperature gradient) with the process benchmark thresholds preset in the energy conversion rule base, define the weights of the energy output index and the carbon emission intensity index, calculate the multi - objective reward value through weighted summation, dynamically evaluate the comprehensive benefits of the current energy conversion path, and ensure the coordinated optimization of the sorting and energy conversion objectives.

[0064] Among them, the process benchmark thresholds preset in the energy conversion rule library are obtained through long-term industrial operation data statistics. Analyze the historical operation records of the anaerobic reactor and the biodiesel reactor, and extract the typical operating condition parameter characteristics (such as methane production rate of 15 m³ / ton, biodiesel purity of 95%, and catalyst activity of 0.8). Based on the multi-objective optimization algorithm, calculate the optimal balance point of energy output and carbon emissions, set the calculation result as the process benchmark threshold, and store it in the rule library. During the operation process, the dynamic update and maintenance of the process benchmark threshold are realized through real-time monitoring data and the federated learning model to ensure the continuous optimization of process parameters.

[0065] S3.3: Based on the reinforcement learning state space and the multi-objective reward function, execute the deep deterministic policy gradient algorithm optimization, and output the range of anaerobic digestion temperature adjustment, the increment of stirring frequency, and the compensation coefficient of the microbial community dosage.

[0066] For example, in the optimization of the anaerobic digestion process, when the reinforcement learning state space monitors that the current methane production rate is 12.5 m³ / ton (lower than the benchmark threshold of 15 m³ / ton), the biodiesel purity is 92% (lower than the benchmark threshold of 95%), the catalyst activity coefficient is 0.85 (within the normal range), and the reactor temperature gradient is 1.2 °C / h (within the allowable range), the deep deterministic policy gradient algorithm outputs the optimization parameters through the following steps.

[0067] Specifically, during the optimization process of the anaerobic digestion process, the policy network (Actor) of the deep deterministic policy gradient algorithm outputs specific control parameters: the anaerobic digestion temperature adjustment is +2.3 °C (raising the original temperature of 35 °C to 37.3 °C), the stirring frequency increases by 1.5 Hz (adjusting from 10 Hz to 11.5 Hz), and the compensation coefficient of the microbial community dosage is set to 1.1 (correcting the original dosage of 100 kg to 110 kg); At the same time, the value network (Critic) evaluates the current action value R = 0.6×12.5 + 0.4×92 - 0.3×1.2 - 0.7×0.85 = 17.35 based on the multi-objective reward function, and predicts that the reward value can be increased to 19.8 (approaching the theoretical maximum value of 22.5) after executing the new parameters; the monitoring data after actual execution shows that the methane production rate has increased to 14.2 m³ / ton (close to the benchmark threshold of 15 m³ / ton), and the biodiesel purity has improved to 94% (still slightly lower than the benchmark threshold of 95%, and the weight adjustment mechanism will be triggered subsequently).

[0068] Among them, the deep deterministic policy gradient algorithm refers to a reinforcement learning method based on the Actor-Critic framework. It directly outputs deterministic control parameters (such as temperature adjustment amount, stirring frequency, etc.) in the continuous action space through the policy network (Actor), and at the same time uses the value network (Critic) to evaluate the quality of actions, and realizes the progressive optimization of process parameters through policy gradient update.

[0069] S3.4: Dynamically adjust the weight ratio of energy output to carbon emissions in the multi-objective reward function, and update the parameter mapping relationship in the energy conversion rule base to generate a process parameter instruction set.

[0070] Specifically, dynamically adjust the weight ratio of energy output to carbon emissions in the multi-objective reward function according to the real-time process deviation. When the energy conversion efficiency decreases, increase the energy output weight and decrease the carbon emission weight; synchronously update the process parameter mapping relationship corresponding to the high-fat and high-fiber waste streams in the energy conversion rule base; generate a process instruction set containing the latest control parameters based on the adjusted weights and parameter mapping, and send it to the energy conversion equipment for execution. Among them, the high-fat waste stream triggers the transesterification reaction to generate biodiesel, and the high-fiber waste stream triggers the two-phase anaerobic digestion to generate biogas, and synchronize the updated process instruction set to the federated learning model for subsequent optimization.

[0071] For example, when the methane production rate continuously drops below the benchmark threshold, increase the energy output weight coefficient by 5% and decrease the carbon emission weight coefficient by 5% at the same time; synchronously update the key parameter mapping in the energy conversion rule base, including increasing the catalyst concentration of the high-fat waste stream by 0.2 mol / L and decreasing the pH threshold of the high-fiber waste stream by 0.2; finally generate an instruction set containing the latest process parameters (such as catalyst concentration, pH value, etc.) and send it to the energy conversion equipment for execution.

[0072] It should be noted that by clarifying the conversion paths of "high-fat waste stream → transesterification reaction → biodiesel" and "high-fiber waste stream → two-phase anaerobic digestion → biogas", directly associate the generated sorting instructions with the process parameter instructions, so as to further ensure that the sorting results can accurately match the subsequent energy conversion process requirements.

[0073] S4: Send the process parameter instruction set to the energy conversion equipment, control the temperature gradient, stirring frequency and microbial population addition amount of the anaerobic reactor, and adjust the catalyst concentration and reaction duration of the biodiesel synthesis reactor, and synchronously monitor the methane production rate and biodiesel purity in real time.

[0074] Specifically, it includes the following steps: S4.1: Analyze the process parameter instruction set and verify the control interface protocol of the energy conversion equipment.

[0075] Among them, the energy conversion equipment includes an anaerobic reactor and a biodiesel synthesis reactor.

[0076] Specifically, analyze the target value of the temperature gradient, the increment of the stirring frequency, and the compensation coefficient of the microbial inoculum dosage in the process parameter instruction set, and at the same time extract the catalyst concentration and reaction duration configuration parameters of the biodiesel synthesis reactor; through the equipment control interface protocol, match the communication protocol and data format compatibility of the energy conversion equipment, and complete the protocol handshake and parameter verification before the execution of the process parameter instruction set.

[0077] S4.2: Hierarchically control the coordinated operation of the anaerobic reactor and the biodiesel reactor.

[0078] Specifically, based on the verified process parameter instruction set, hierarchically control the anaerobic reactor and the biodiesel reactor. Anaerobic reactor control layer: Execute heating / cooling control according to the target value of the temperature gradient, dynamically adjust the stirring frequency to achieve uniform mixing of materials, and calibrate the microbial inoculum feeding rate in real time according to the compensation coefficient of the microbial inoculum dosage; Biodiesel reactor control layer: Inject the catalyst solution according to the catalyst concentration configuration gradient, and synchronously adjust the transesterification reaction duration to match the oil conversion efficiency; Coordination control layer: Through real-time regulation, ensure the dynamic matching of the process parameters of the two-stage reactor, such as the dynamic matching of the catalyst injection rate and the temperature gradient change, and its precise coordination to keep the reaction conditions in the best synchronization.

[0079] S4.3: Continuously monitor the methane production rate through the anaerobic reactor, track the biodiesel purity using the biodiesel synthesis reactor, and synchronously collect the catalyst activity coefficient and temperature gradient deviation data.

[0080] Generate a multi-dimensional quality assessment report including energy conversion efficiency, carbon emission intensity, and equipment health status based on the monitoring data, and upload it to the digital thread platform in real time for full-process tracking.

[0081] Among them, the monitoring data includes the methane production rate, biodiesel purity, catalyst activity coefficient, and temperature gradient deviation collected in real time.

[0082] S5: Track the energy output data and carbon emission intensity of the whole process through the digital thread, dynamically correct the weight coefficients of the federated learning model, and update the energy conversion rule base.

[0083] Specifically, it includes the following steps: S5.1: Integrate the initial component data, dynamic sorting instruction set execution log, process parameter instruction set configuration, and real-time monitoring data through the digital thread platform to construct a full-link digital twin covering waste sorting to energy conversion.

[0084] Specifically, by integrating in real time the initial composition data collected by the hyperspectral sorter, the execution log of the dynamic sorting instruction set generated by the multi-stage linkage equipment, the process parameter configuration received by the energy conversion equipment, and the real-time monitoring data of the sensors through the digital thread platform, a multi-source heterogeneous data is uniformly encoded into a three-dimensional space-time series model by using a spatio-temporal alignment algorithm. This three-dimensional space-time series model establishes a real-time association between the spatial distribution of waste components and the operating state of the equipment through dynamic coordinate mapping technology, and integrates an energy efficiency-carbon emission collaborative analysis algorithm to construct a full-link digital twin covering from waste sorting to energy conversion.

[0085] It should be noted that the energy efficiency-carbon emission collaborative analysis algorithm refers to calculating energy efficiency by integrating the existing anaerobic digestion ADM1 model and the kinetic equation of transesterification reaction, combining the carbon accounting methods of the IPCC emission factor method and the life cycle assessment database (such as Ecoinvent), and using a multi-objective optimization algorithm (such as NSGA-II) to find the balance point of maximizing energy efficiency and minimizing carbon emission in the Pareto optimal solution set.

[0086] The spatio-temporal alignment algorithm is used to map multi-source heterogeneous data to a unified timestamp and spatial coordinate system to generate a traceable energy efficiency-carbon emission correlation dataset.

[0087] S5.2: Using the energy efficiency-carbon emission correlation dataset as input, a multi-objective optimization algorithm is used to dynamically correct the weight coefficients of the federated learning model.

[0088] It should be noted that the multi-objective optimization algorithm takes maximizing energy efficiency (methane production rate + biodiesel production) and minimizing carbon emission (unit energy consumption + catalyst loss) as the dual optimization objectives, and determines the optimal weight combination through Pareto front analysis.

[0089] Specifically, the spatio-temporal covariance weight of the high-precision prediction area (such as the methane production rate prediction error < 5%) is increased, and the parameter aggregation weight of the area with large historical data fluctuations is reduced; the corrected weight coefficients are encrypted and transmitted back to each regional federated node through the differential privacy mechanism to drive the global model optimization in the next training cycle.

[0090] S5.3: According to the process optimization trajectory and abnormal event characteristics recorded by the digital thread, the incremental update of the energy conversion rule library is realized by defining three types of operations: new rules, elimination rules, and correction rules.

[0091] Specifically, new rules: Generate new mapping relationships based on parameter combinations that have successfully improved energy conversion efficiency (such as specific temperature-catalyst concentration ratios); elimination rules: Mark historical parameter configurations that have not been triggered for three consecutive cycles as invalid rules and remove them from the rule library; correction rules: Adjust the threshold conditions for the conversion path of high-oil / high-fiber waste streams (such as increasing the oil content threshold from 35% to 38%).

[0092] This embodiment also provides a computer device applicable to the case of the energy conversion treatment method for mixed food waste, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the energy conversion treatment method for mixed food waste proposed in the above embodiment.

[0093] This computer device may be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner. The wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0094] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the energy conversion treatment method for mixed food waste proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0095] In summary, by combining the federated learning model with hyperspectral sorting technology, the present invention accurately captures the spatio-temporal distribution characteristics of the components of mixed food waste, improving the sorting accuracy and energy conversion efficiency. By integrating multi-region encrypted component data with a differential privacy perturbation mechanism, a global spatio-temporal tensor prediction model is constructed to enhance the robustness of dynamic prediction of waste components and data privacy security. By adopting multi-objective carbon emission constraint conditions in the reinforcement learning decision engine, the optimization of the energy conversion process is deeply coupled with the carbon neutrality goal, enhancing the environmental friendliness and energy efficiency synergy of the treatment process, and realizing the efficient and low-carbon closed-loop operation of the energy treatment of food waste.

[0096] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for energy conversion treatment of mixed food waste, characterized in that: including collecting the initial composition data of mixed food waste, encrypting it and uploading the encrypted data of the region to the federated learning model training platform, aggregating the encrypted data of multiple regions and training the global garbage composition prediction model, and outputting the composition distribution probability matrix of mixed food waste in the future period; inputting the composition distribution probability matrix into a hyperspectral sorter, and using 3D-CNN to analyze the spatial distribution of organic matter, oil and impurities in the unsorted mixed waste stream in real time to generate a dynamic sorting instruction set; controlling the multi-stage linkage equipment according to the dynamic sorting instruction set, outputting a high-oil garbage stream and a high-fiber garbage stream with homogeneous composition, matching the preset energy conversion rule base through the reinforcement learning decision-making engine, dynamically selecting the optimal energy conversion path and generating a process parameter instruction set; sending the process parameter instruction set to the energy conversion equipment, controlling the temperature gradient, stirring frequency and bacterial inoculation amount of the anaerobic reactor, and adjusting the catalyst concentration and reaction duration of the biodiesel synthesis reactor, and synchronously monitoring the methane production rate and biodiesel purity in real time; tracking the energy output data and carbon emission intensity of the whole process through the digital thread, dynamically correcting the weight coefficient of the federated learning model and updating the energy conversion rule base.

2. The energy conversion treatment method for mixed food waste as described in claim 1, wherein: The output of the composition distribution probability matrix of mixed food waste in the future period includes the following steps In the federated learning model training platform, perform homomorphic decryption on the encrypted data of each region, add differential privacy noise to the decrypted data, and extract local model parameters; aggregate local model parameters through a dynamic weight allocation algorithm, and jointly determine the weights of local model parameters by the historical variation coefficient and seasonal fluctuation amplitude value of the regional garbage composition; Based on the weights of local model parameters, construct a spatio-temporal feature fusion space through an attention mechanism, and integrate the time series gradient and spatial covariance matrix of the garbage composition in each region into a global spatio-temporal tensor; input the global spatio-temporal tensor into a Bayesian probability network to generate a composition distribution probability matrix.

3. The energy conversion treatment method for mixed food waste as claimed in claim 2, wherein: The generation of the dynamic sorting instruction set includes the following steps collecting multi-band reflectance spectral data of the unsorted mixed waste stream and constructing a three-dimensional hyperspectral tensor; The unsorted mixed waste stream refers to the physical material that enters the hyperspectral sorter in real time through the conveyor belt; using the depthwise separable convolutional layer of 3D-CNN to extract the spatial distribution characteristics of organic matter, oil and impurities, and synchronously calculating the temporal change rate of the spectral absorption peak; fusing the spatial distribution characteristics and the temporal change rate of the spectral absorption peak, predicting the sorting lag compensation amount through a gated recurrent unit, and generating a three-dimensional dynamic sorting instruction set.

4. The method for energy conversion treatment of mixed food waste according to claim 3, characterized in that: The dynamic selection of the optimal energy conversion path and the generation of the process parameter instruction set include the following steps real-time monitoring of the real-time methane production rate, biodiesel purity, catalyst activity coefficient and reactor temperature gradient in the multi-stage linkage equipment as monitoring data, and constructing a reinforcement learning state space; Based on the reinforcement learning state space, construct a multi-objective reward function by combining the energy output index and the carbon emission intensity index in a weighted summation manner; Based on the state space of reinforcement learning and the multi-objective reward function, perform deep deterministic policy gradient algorithm optimization, and output the range of anaerobic digestion temperature adjustment, the increment of stirring frequency, and the compensation coefficient of microbial inoculum dosage. Dynamically adjust the weight ratio of energy output and carbon emission in the multi-objective reward function, and update the parameter mapping relationship in the energy conversion rule base to generate a process parameter instruction set.

5. The method for energy conversion treatment of mixed food waste as claimed in claim 1, wherein: The initial component data includes oil content, moisture ratio, and impurity type.

6. The method for energy conversion treatment of mixed food waste as claimed in claim 1, wherein: The hyperspectral sorter refers to a device integrating hyperspectral imaging, 3D-CNN real-time analysis, and sorting instruction issuance functions.

7. The method for energy conversion treatment of mixed food waste as claimed in claim 1, characterized in that: The control multi-stage linkage device includes adaptively adjusting the rotational speed gradient of the shear crusher, the pressure threshold of the screw dehydrator, and the trigger condition for performing low-temperature oil removal.

8. The method for energy conversion treatment of mixed food waste as claimed in claim 1, characterized in that: The high-oil-content waste stream triggers a transesterification reaction to produce biodiesel, and the high-fiber waste stream triggers two-phase anaerobic digestion to produce biogas.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for energy treatment of mixed food waste according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for energy treatment of mixed food waste according to any one of claims 1 to 8.

Images